Machine learning-driven prediction of intratumoral tertiary lymphoid structures in hepatocellular carcinoma using contrast-enhanced CT imaging and integrated clinical data

Jun WU¹Xiaodong Feng¹Lin Na²Wei Zhang¹Yang Guo¹Ziwei Zhu¹Qiongyuan Ren³Zhifan Zuo⁴Weng Kung Peng^5*Lei Han^1*

• ¹Department of Hepatobiliary Surgery, The General Hospital of Northern Theater Command,, Shenyang, Liaoning Province,, China
• ²Department of Central Laboratory,The First Affiliated Hospital of China Medical University, Shenyang, Liaoning, China
• ³Dalian Medical University, The General Hospital of Northern Theater Command Training Base for Graduate, Shenyang, Liaoning Province, China
• ⁴Gynecological Radiotherapy Ward, Liaoning Provincial Cancer Hospital, Shenyang, Liaoning province, China
• ⁵Institute of Advanced Co-Creation Studies, Center for Quantum Information and Quantum Biology, The University of Osaka, Japan

Purpose: We developed a machine learning framework to predict the presence of tertiary lymphoid structures (TLSs) within tumors in patients with hepatocellular carcinoma (HCC). This framework uses computed tomography (CT) imaging and clinical data collected before surgery, providing a noninvasive method for prediction. Methods: We conducted a retrospective analysis of HCC patients who underwent surgery at the General Hospital of the Northern Theater Command's Hepatobiliary Surgery Department between January 2017 and October 2024. Using Python software, we extracted radiomic features from preoperative CT images (arterial and portal venous phases). We then selected features associated with intratumoral TLSs using statistical methods, including intraclass correlation coefficient (ICC), Pearson correlation, t-tests, and LASSO regression. Three models were developed—clinical, radiomics, and combined—using machine learning techniques and independent clinical predictors. A predictive nomogram was created and evaluated using the area under the ROC curve (AUC) and calibration analysis. Results: Our study included 171 HCC patients, with 80 showing negative and 91 showing positive expression of intratumoral TLSs. Multivariate analysis identified the albumin-bilirubin (ALBI) score as an independent predictor of intratumoral TLSs expression. The combined model demonstrated the highest predictive accuracy, with AUCs of 0.947 in the training set and 0.909 in the validation set, outperforming both the clinical (AUC: 0.709 training, 0.714 validation) and radiomics (AUC: 0.935 training, 0.890 validation) models. Conclusion: Our combined machine learning model, which integrates preoperative CT imaging and clinical data, provides an accurate, noninvasive method for assessing intratumoral TLSs expression in HCC. This tool has the potential to enhance clinical decision-making, guide therapeutic planning, and facilitate personalized treatment strategies for HCC patients.